Estimating NOx emissions with numerical modeling and deep learning estimated complete surface NO2 map
Abstract
To defeat the limitations in estimating NOx emissions based on satellite images (e.g., missing data and contribution of background sources), we applied the complete surface NO2 map estimated by using the deep neural network (DNN) in the summer of 2017 over the contiguous United States (CONUS). Additionally, we leveraged the partial convolutional neural network (PCNN) approach to impute the OMI tropospheric NO2 column density, which is one of the predictor variables used in the DNN model. Based on the ten-fold cross-validation approach, the performance of the DNN model at estimating surface NO2 was evaluated, showing high capability with a strong correspondence (R: 0.92, IOA: 0.96, MAB: 1.52, RMSE: 2.34). Furthermore, we used the Community Multi-scale Air Quality (CMAQ) model at 12 km grid spacing to conduct inverse modeling based on the iterative finite difference mass balance (iFDMB) technique on a daily time resolution. Compared to the prior emissions, the inversion suggested 3.66 times higher NOx emissions over CONUS, significantly mitigating the underestimation of the prior emissions. We also evaluated the NO2 concentrations against U.S. EPA Air Quality System (AQS) measurement, exhibiting an increased correlation coefficient (from 0.55 to 0.68) and decreased RMSE (from 3.07 to 2.77), even though a few stations showed a higher bias due to limited spatial resolution to capture a small peak of NO2 in urban areas. Results showed the primary benefits of incorporating the neural network technique into numerical modeling in estimating emissions, especially where had limited improvements from satellite column measurements.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.A45I1955J